NVIDIA B300
The B300 is NVIDIA's Blackwell Ultra data center GPU, a binned and re-engineered Blackwell die that trades the B200's 8-Hi HBM3e stacks for 12-Hi stacks to land 288 GB per GPU, and raises the power ceiling to 1,400 W to push dense NVFP4 throughput about 1.5x past B200. It ships on the same 8-GPU HGX baseboard pattern as B200 and is also the GPU building block of the GB300 NVL72 rack (72 GPUs, 36 Grace CPUs, fully liquid-cooled). In a private AI build it is the part to reach for when a single GPU's memory ceiling, not raw FLOPS, is what is blocking a large-context or reasoning-model deployment. As of mid-2026 it is NVIDIA's current top-of-stack Blackwell part, shipping since January 2026, and new OEM quotes are increasingly steering toward it over B200.
- The FP4 sparse-to-dense ratio on B300 is only about 1.33x (15 to 20 PFLOPS), not the 2x you'd expect from structured sparsity and that B200 actually delivers (9 to 18 PFLOPS). Do not assume Blackwell Ultra scales sparsity the same way standard Blackwell does when you're modeling throughput.
- 1,400 W per GPU with no air-cooled option means the facility conversation (liquid loop, CDU capacity, floor loading) has to happen before you place the GPU order, not after. This is a harder requirement than B200, where air-cooled HGX chassis exist.
- The jump to 288 GB matters most when a single model replica, its KV cache, or its optimizer state doesn't fit in 180 GB. If your working set fits comfortably on B200, the memory headroom on B300 buys you little beyond the ~1.5x dense FP4 uplift.
- H200 (141 GB HBM3e, air-cooled, Hopper software stack) is still the pragmatic choice for teams that need memory-bandwidth-bound inference today and cannot absorb a liquid-cooling retrofit for B300's 1,400 W ceiling.
- Treat published NVFP4 throughput as a ceiling, not a delivered number: it assumes a quantization pipeline that actually produces valid NVFP4 weights and a kernel stack tuned for it. Benchmark your own model before sizing a cluster on the PFLOPS line.
How much does an NVIDIA B300 cost?
Reported figures put a single B300 GPU around $50,000, with a fully configured 8-GPU DGX B300 system in the $300,000 to $500,000 range. A full GB300 NVL72 rack has been estimated near $3.7 to $4.0 million based on large hyperscaler order commentary. These are reported market figures, not published NVIDIA list prices.
B200 vs B300: what's the actual difference?
B300 has 288 GB of HBM3e versus B200's 180 GB, a 1,400 W power ceiling versus B200's 1,000 W, and roughly 1.5x B200's dense NVFP4 throughput per GPU. B300 also requires direct liquid cooling in every deployment path, while B200 has air-cooled HGX options.
Does B300 require liquid cooling?
Yes. At up to 1,400 W per GPU there is no air-cooled HGX B300 or DGX B300 SKU; every reference platform (HGX B300, DGX B300, GB300 NVL72) ships as direct-liquid-cooled.
What is GB300 NVL72?
GB300 NVL72 is NVIDIA's rack-scale system built from B300: 72 Blackwell Ultra GPUs and 36 Grace CPUs in one liquid-cooled rack, connected by a 130 TB/s aggregate NVLink fabric, delivering roughly 1.1 exaFLOPS of dense FP4 throughput at the rack level.